Abstract: Gearbox fault data are usually noisy, multicomponent, and nonstationary. As a result, the classification of complex gearbox vibration data with similar fault patterns is still one of intractable problems for gearbox fault diagnosis. Detrended fluctuation analysis (DFA) is utilized to analyze the scaling behavior of gearbox vibration data and the scaling exponent will have an abrupt change with the gradual increase in time scales. Thus, a two-dimensional vector containing two scaling exponent carries definite physical meaning and can be used as feature parameters to describe the underlying dynamic mechanism hiding in gearbox vibration data. Consequently, a novel method for gearbox fault diagnosis based on the double-scaling-exponent characteristic of nonstationary time series is proposed. Moreover, the proposed method, as well as Fourier transform, wavelet transform and the single-scaling-exponent method, is exploited to classify the normal, slight-worn, medium-worn and broken-gear vibration data from a four-speed motorcycle gearbox. The results show that the proposed method is sensitive to the subtle differences between two similar fault patterns and capable of solving the classification of two complex gearbox vibration data with similar fault patterns, which another three methods fail to do.

Key words: Detrended fluctuation analysis, Fault diagnosis, Gearbox, Scaling exponent